+/* Copyright (c) 2013-2014. The SimGrid Team.
+ * All rights reserved. */
+
+/* This program is free software; you can redistribute it and/or modify it
+ * under the terms of the license (GNU LGPL) which comes with this package. */
+
#include "colls_private.h"
-#ifndef NUM_CORE
-#define NUM_CORE 8
-#endif
+
int bcast_SMP_binary_segment_byte = 8192;
rank = smpi_comm_rank(comm);
size = smpi_comm_size(comm);
-
- if(size%NUM_CORE)
- THROWF(arg_error,0, "bcast SMP binary can't be used with non multiple of NUM_CORE=%d number of processes ! ",NUM_CORE);
+ if(smpi_comm_get_leaders_comm(comm)==MPI_COMM_NULL){
+ smpi_comm_init_smp(comm);
+ }
+ int host_num_core=1;
+ if (smpi_comm_is_uniform(comm)){
+ host_num_core = smpi_comm_size(smpi_comm_get_intra_comm(comm));
+ }else{
+ //implementation buggy in this case
+ return smpi_coll_tuned_bcast_mpich( buf , count, datatype,
+ root, comm);
+ }
int segment = bcast_SMP_binary_segment_byte / extent;
int pipe_length = count / segment;
int remainder = count % segment;
- int to_intra_left = (rank / NUM_CORE) * NUM_CORE + (rank % NUM_CORE) * 2 + 1;
- int to_intra_right = (rank / NUM_CORE) * NUM_CORE + (rank % NUM_CORE) * 2 + 2;
- int to_inter_left = ((rank / NUM_CORE) * 2 + 1) * NUM_CORE;
- int to_inter_right = ((rank / NUM_CORE) * 2 + 2) * NUM_CORE;
- int from_inter = (((rank / NUM_CORE) - 1) / 2) * NUM_CORE;
- int from_intra = (rank / NUM_CORE) * NUM_CORE + ((rank % NUM_CORE) - 1) / 2;
+ int to_intra_left = (rank / host_num_core) * host_num_core + (rank % host_num_core) * 2 + 1;
+ int to_intra_right = (rank / host_num_core) * host_num_core + (rank % host_num_core) * 2 + 2;
+ int to_inter_left = ((rank / host_num_core) * 2 + 1) * host_num_core;
+ int to_inter_right = ((rank / host_num_core) * 2 + 2) * host_num_core;
+ int from_inter = (((rank / host_num_core) - 1) / 2) * host_num_core;
+ int from_intra = (rank / host_num_core) * host_num_core + ((rank % host_num_core) - 1) / 2;
int increment = segment * extent;
- int base = (rank / NUM_CORE) * NUM_CORE;
- int num_core = NUM_CORE;
- if (((rank / NUM_CORE) * NUM_CORE) == ((size / NUM_CORE) * NUM_CORE))
- num_core = size - (rank / NUM_CORE) * NUM_CORE;
+ int base = (rank / host_num_core) * host_num_core;
+ int num_core = host_num_core;
+ if (((rank / host_num_core) * host_num_core) == ((size / host_num_core) * host_num_core))
+ num_core = size - (rank / host_num_core) * host_num_core;
// if root is not zero send to rank zero first
if (root != 0) {
// when a message is smaller than a block size => no pipeline
if (count <= segment) {
// case ROOT-of-each-SMP
- if (rank % NUM_CORE == 0) {
+ if (rank % host_num_core == 0) {
// case ROOT
if (rank == 0) {
//printf("node %d left %d right %d\n",rank,to_inter_left,to_inter_right);
(MPI_Status *) xbt_malloc((size + pipe_length) * sizeof(MPI_Status));
// case ROOT-of-each-SMP
- if (rank % NUM_CORE == 0) {
+ if (rank % host_num_core == 0) {
// case ROOT
if (rank == 0) {
for (i = 0; i < pipe_length; i++) {